A Novel Acceleration Method for DGTD Algorithm on Sunway TaihuLight

Abstract

In this paper, a novel acceleration method for dis-continuous Galerkin time domain (DGTD) algorithm on Sunway platform is proposed with a multi-level speedup technique. In the proposed method, the message passing interface (MPI) is used to connect each core-group (CG) and the registor level massaging passing interface (RL-MPI) is used to connect the slave core on each computing processing element (CPE) cluster, which efficiently overcomes the bottleneck caused by discrete memory accessing. Numerical results show that the performance of DGTD on SW26010 CPU can be dramatically improved by using RL-MPI.